In KALDI, Can I specify an utterance and run MLLR using an existing model, forcing all utterances in a training-data set to align to the utterance I specify and outputting a likelihood (or some probabilistic measure) that each utterance is a match to the utterance specified? [Some call this 'forced alignment'].
I have looked into the steps/align_*.sh scripts and am not sure if they are where I should be looking. Thanks ^_^
-Daniel
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I think you may not understand what MLLR is.
You could do what you say by creating a data directory where you
replace all the "text" files with a given utterance, but even if you
created the likelihoods (I think it may be an optional output of
gmm-align-compiled) it wouldn't give you very meaningful measure at
all. Better to run data/cleanup/find_bad_utts.sh which I think gives
more meaningful diagnostics of whether an utterance's transcription is
correct.
Dan
In KALDI, Can I specify an utterance and run MLLR using an existing model,
forcing all utterances in a training-data set to align to the utterance I
specify and outputting a likelihood (or some probabilistic measure) that
each utterance is a match to the utterance specified? [Some call this
'forced alignment'].
I have looked into the steps/align_*.sh scripts and am not sure if they are
where I should be looking. Thanks ^_^
In KALDI, Can I specify an utterance and run MLLR using an existing model, forcing all utterances in a training-data set to align to the utterance I specify and outputting a likelihood (or some probabilistic measure) that each utterance is a match to the utterance specified? [Some call this 'forced alignment'].
I have looked into the steps/align_*.sh scripts and am not sure if they are where I should be looking. Thanks ^_^
-Daniel
I think you may not understand what MLLR is.
You could do what you say by creating a data directory where you
replace all the "text" files with a given utterance, but even if you
created the likelihoods (I think it may be an optional output of
gmm-align-compiled) it wouldn't give you very meaningful measure at
all. Better to run data/cleanup/find_bad_utts.sh which I think gives
more meaningful diagnostics of whether an utterance's transcription is
correct.
Dan
On Tue, Jul 7, 2015 at 11:01 AM, Daniel Douglas daniel-tee@users.sf.net wrote: